Do 'Expected' goals statistics suck?

[Lafleurs Guy]

I pretty much hate all the"advanced statistics". They're absolutely meaningless. Way too many variables in hockey.

Imo

Hockey prediction is closer to pinball than baseball. Much closer.

Statistical comparison and probabilistic outcomes are not something many of our brains like but the puck is either on edge or flat when the one timer comes through, the goalie perhaps can’t even see that an opponent is shooting, the puck hits a guy in the pants, then a skate and goes in 5 hole.

This is pure chaos.

Compare this to: guy throws his 52nd fastball and guy swings.

One is way easier to model than the other, much more predicable with results that are way easier to measure even in weird ways.

What advanced stats do and are is just a handful of measuring sticks. They attempt to place the same measuring stick on every team, player, and event. They can’t help their job is to define chaos.

The root of your problem is ice hockeys chaotic nature, not the evaluation of it.

Understanding their fundamental source will allow you to recognize them for what they are.

I’d suggest that they’re a great tool that joins others to help describe historic events and predict future ones.

Corsi and Fenwick numbers have been shown to be the most accurate predictors of cup winners. It’s not ‘pinball’ or random.

Those stats give you a really good idea of where the play is most of the time. And it’s only common sense that the team that soends most of its time in the opposing zone has a better chance of winning. Nothing magical about that.

All that being said - that doesn’t mean the ‘win-O-meter is going to be valuable for anything on any given game.

 

[Dirty Dog]

They certainly aren’t good. But they are better than the horrible expected saves stats…those are truly bad

 

[Dr Pepper]

"Expected goals" or any other expected stat has never made any sense to me.

 

[Machinehead]

"Expected goals" or any other expected stat has never made any sense to me.

You have a shot attempt. Based on years of data of how often similar shots are goals, you apply a percentage chance of that shot going in. You convert that percentage to a decimal. That's how many expected goals you get.

There's really not that much to it and it's not particularly voodoo.

 

[Dr Pepper]

You have a shot attempt. Based on years of data of how often similar shots are goals, you apply a percentage chance of that shot going in. You convert that percentage to a decimal. That's how many expected goals you get.

There's really not that much to it and it's not particularly voodoo.

Why not just use the statistic we already have for such events, namely a goalie's save percentage?

Number of shots stopped, divided by number of shots faced.

Simple. Don't need "years of data" to create these new stats for what "might" happen.

 

[Machinehead]

Why not just use the statistic we already have for such events, namely a goalie's save percentage?

Number of shots stopped, divided by number of shots faced.

Simple. Don't need "years of data" to create these new stats for what "might" happen.

Because it isn't a goaltending stat?

 

[Beezeral]

Why not just use the statistic we already have for such events, namely a goalie's save percentage?

Number of shots stopped, divided by number of shots faced.

Simple. Don't need "years of data" to create these new stats for what "might" happen.

Because stats have evolved? A goalie stopping 21/22 low danger shots in a low event game is going to finish a game with a higher save percentage than the goalie who gave up 3 goals on 17 shots but 10 of them were premium chances.

Which would you say played the better game?

 

[the valiant effort]

The real analytics behind closed doors are focused on load management, maximizing output and just winning the war of attrition.

 

[Dr Pepper]

Because it isn't a goaltending stat?

It kind of is though, isn't it?

If a goalie enters the game with a .914 save percentage on the season, then the "expected goals" should be 8.6% of the total shot attempts, should it not?

Because stats have evolved? A goalie stopping 21/22 low danger shots in a low event game is going to finish a game with a higher save percentage than the goalie who gave up 3 goals on 17 shots but 10 of them were premium chances.

Which would you say played the better game?

Still the first one, obviously, and hopefully when it comes time to figure out who wins the Vezina we don't have voters sitting around a table ignoring real stats and trying to figure out which goalie stopped the most "premium chances" instead.

 

[Machinehead]

It kind of is though, isn't it?

If a goalie enters the game with a .914 save percentage on the season, then the "expected goals" should be 8.6% of the total shot attempts, should it not?

It's tangentially related to goaltending (like everything else is) but nobody uses it that way. The purpose of xG is to understand which teams had the better scoring chances.

Raw save percentage is still in popular use, and not too many people have pushed back on that.

 

[GrumpyKoala]

Change it to score adjusted and the math completely flips. You used one metric and didn’t look at the rest. 50+%. Of the Rangers chances came in the 3rd when the game was no longer in doubt.View attachment 990872

There was 34 minutes of penalty in that game,
I am just guessing from your screenshot but it seem that the deserve to win-o-meter take all situation into account and the second graph do not.

 

[vaspa]

It's tangentially related to goaltending (like everything else is) but nobody uses it that way. The purpose of xG is to understand which teams had the better scoring chances.

Raw save percentage is still in popular use, and not too many people have pushed back on that.

I think some people for whatever reason don't like the stat because expecting and assumedly deserving something that didn't manifest itself on the scoreboard ruffles them the wrong way. I guess the old school discourse full of emphasizing winning and the mystique of winners and clutchness and so on doesn't fit with predicative numbers.

I think If the stat would be called like quality of scoring chances created or something it would be a lot less contentious.

 

[robertmac43]

I pretty much hate all the"advanced statistics". They're absolutely meaningless. Way too many variables in hockey.

Yeah its far less effective at telling the game story then Baseball or even soccer.

 

[Fedz]

I've always felt that CF% gives a much clearer picture of who controlled play at evens.

 

[JoVel]

I mostly hate the entire discourse in advanced stats in hockey, or sports in general, because it always feels like the two parties are sports people who don't understand anything about statistics, and statisticians who don't understand anything about sports.

 

[eojsmada]

For me, the problem with all of the advanced metrics is that they are all relative to the strengths of both the shooter and the goalie. If a player shoots from the left dot in a PP (e.g. ovechkin's house), the shot taken is variable to the ability of both the shooter able to accurately shoot the puck as well as the goalie's ability to both track and stop the puck. So while, on paper, it seems like advanced stats "should" be useful, they really are only useful on a micro level. One could argue that roster construction or team "systems" can be enhanced by these micro-level statistics, but a coach would be sent into a dizzying spin trying to accurately figure out how to deploy a roster both before and during a game.

 

[Machinehead]

For me, the problem with all of the advanced metrics is that they are all relative to the strengths of both the shooter and the goalie. If a player shoots from the left dot in a PP (e.g. ovechkin's house), the shot taken is variable to the ability of both the shooter able to accurately shoot the puck as well as the goalie's ability to both track and stop the puck. So while, on paper, it seems like advanced stats "should" be useful, they really are only useful on a micro level. One could argue that roster construction or team "systems" can be enhanced by these micro-level statistics, but a coach would be sent into a dizzying spin trying to accurately figure out how to deploy a roster both before and during a game.

If a guy takes a shot and gets a given number of xG, and then Ovechkin takes that same shot, it's gonna be the same number. That's a valid criticism.

That's why when assessing teams' chances going into the playoffs, I look at things like xG but I also look at the lineup on paper to see who has the type of talent that will over/under perform those numbers. The 2018 Capitals weren't a great xG team but they had a guy that shoots 13% with volume.

But that's something you have to take into account, not necessarily throw the baby out with the bath water.

 

[bucks_oil]

I feel like everytime I go under the hood and really look at them, they just do not represent reality.

For example, Moneypuck had this Justin Danforth goal as .16 XG, which to my understanding is them saying this goal should only go in 16/100 times on an average NHL goalie (please correct me if I am wrong). IMO, it's closer to 96/100 than it is 16.




Idk who watched the Rangers/Jackets tonight but it's absolutely laughable that these stats 'say' the Rangers should have won that game most of the time. They were absolutely dreadful defensively and in pretty much all phases of the game.

I feel like I laugh my ass off at the 'Win O Meter' more than half the time they get posted in GDTs/PGT's (again as being wrong and unreflective of reality)

Help me. Am I misinterpreting the information? Should these stats and source (Moneypuck) be thrown out the window? Is there a better source (where you can actually pinpoint specific shots/goals and see their 'expectedness')

Yes... the public models suck and you gave a PERFECT example. They DO NOT account for east-west puck movement prior to a shot.

I'm told some of the proprietary models are working on this kinda stuff, but until they do... I find them total BS.

You are better off with ACTUAL goal stats and matching with eye test. It may be unfair to attribute blame/credit with ACTUAL stats, but at least you are starting with something that was REAL and thus it's worth evaluating.

I'm not sure its right to statistically penalize a defender who manages to consistently angle an attacker toward the face-off dot where he fires off an off-balance wrister while under pressure... if he makes that happen 10 times a game and the offense-only defender on his same team gives up two cross-ice one-timers on 2-1 breaks from the same spot... the offense-only guy is going to have a better xGA than the guy who was actually playing defense properly.

 

[Romang67]

It's just a statistical distribution. It's not magic. People were complaining about shot metrics' not taking shot location into account. Now the shot metrics take shot location into account. Pre-shot puck movement is available in tracking data, and if not already taken into account, should and will be taken into account.

It's better than what we had, and it is improving.

If a guy takes a shot and gets a given number of xG, and then Ovechkin takes that same shot, it's gonna be the same number. That's a valid criticism.

That's why when assessing teams' chances going into the playoffs, I look at things like xG but I also look at the lineup on paper to see who has the type of talent that will over/under perform those numbers. The 2018 Capitals weren't a great xG team but they had a guy that shoots 13% with volume.

But that's something you have to take into account, not necessarily throw the baby out with the bath water.

Is that a valid criticism though? That feels more like misunderstanding what xG is, or just generally what expected value in statistics means.

 

[eojsmada]

Yes... the public models suck and you gave a PERFECT example. They DO NOT account for east-west puck movement prior to a shot.

I'm told some of the proprietary models are working on this kinda stuff, but until they do... I find them total BS.

You are better off with ACTUAL goal stats and matching with eye test. It may be unfair to attribute blame/credit with ACTUAL stats, but at least you are starting with something that was REAL and thus it's worth evaluating.

I'm not sure its right to statistically penalize a defender who manages to consistently angle an attacker toward the face-off dot where he fires off an off-balance wrister while under pressure... if he makes that happen 10 times a game and the offense-only defender on his same team gives up two cross-ice one-timers on 2-1 breaks from the same spot... the offense-only guy is going to have a better xGA than the guy who was actually playing defense properly.

Agree. There are too many variables and legitimate tactics involved to assign every shot taken a generalized statistical value. And then include the various abilities of the goaltender as another layer of variables, combined with shift-length, time of the game, endurance levels of individual players, and a plethora of other things to be considered. It's just all too much to take a particular data point and have it count as a meaningful statistic on a macro level.

 

[Machinehead]

It's just a statistical distribution. It's not magic. People were complaining about shot metrics' not taking shot location into account. Now the shot metrics take shot location into account. Pre-shot puck movement is available in tracking data, and if not already taken into account, should and will be taken into account.

It's better than what we had, and it is improving.

Is that a valid criticism though? That feels more like misunderstanding what xG is, or just generally what expected value in statistics means.

I think it's valid insofar as, some players will outperform their chances consistently and it's not gonna show up. I think that's a valid "flaw" if you want to call it that.

I also don't think that needs to be corrected, necessarily. Like, "this is now a 3% better chance because Auston Matthews took the shot" feels subjective, and the elite players in this league are already factoring in to how many shots go in from that area.

I just think it's fair to say, in a small sample, "well, that was Ovechkin, so..." That's a point I'm willing to consider.

I think the weight on that point is outsized. Like I said, the shot location data is already accounting for a league with elite players in it, and yeah, if one team has a lot of talent you should already just take that into account. People like to criticize analytics by yelling "context!" and don't want a real discussion on context. They want to ignore the stats because of one thing they don't like.

 

[eojsmada]

I think it's valid insofar as, some players will outperform their chances consistently and it's not gonna show up. I think that's a valid "flaw" if you want to call it that.

I also don't think that needs to be corrected, necessarily. Like, "this is now a 3% better chance because Auston Matthews took the shot" feels subjective, and the elite players in this league are already factoring in to how many shots go in from that area.

I just think it's fair to say, in a small sample, "well, that was Ovechkin, so..." That's a point I'm willing to consider.

I think the weight on that point is outsized. Like I said, the shot location data is already accounting for a league with elite players in it, and yeah, if one team has a lot of talent you should already just take that into account. People like to criticize analytics by yelling "context" and don't want a real discussion on context. They want to ignore the stats because of one thing they don't like.

For me, the problem lies in thati t's just a series of tracked data sets that someone decides, on a whim, to assign a higher value to because they may or may not have a theory as to how goals are scored or stopped. The problem is that there is no conclusion to the data, just the data. Which is why it seems devoid of any context. At least that's how I see it.

Always willing to have a discussion but it has to start with why is it valued over any other statistic, and can you show any level of if/then causality attributable to said data point(s).

 

[Thallis]

What better predicts how many goals per game a team will score next month - the number of goals per game it scored this month or the xG per game it put up this month?

Or what about year to year?

If it predicts better than the current goal pace then it is already very valuable. I'd like the xG sucks crowd in here to consider that.

The "deserve to win'o meter" is kind of tongue in cheek. You're not supposed to take it too seriously.


----------------

On the Danforth goal (holy Werenski!) there is a cross crease pass right before the shot (royal road pass). At one point there were analysts trying to figure out how to build the pre-shot movement into the model, so that, for example, a shot from that spot with a goalie set is like .08 xG while a shot after a cross crease pass is a .4 xG. I would assume that that effort didn't end up in the public models like moneypuck - perhaps pre-shot movement is now built in to the models that teams are using.

Public models like moneypuck do not, but private models like Sportlogiq and ClearSight do. This creates a bias towards volume instead of quality and js the biggest flaw in public models right now and has been for a few years. That's why teams like Carolina are always at the top, while Vegas always seems to be lower than expected.

 

[a mangy Meowth]

Why not just use the statistic we already have for such events, namely a goalie's save percentage?

Number of shots stopped, divided by number of shots faced.

Simple. Don't need "years of data" to create these new stats for what "might" happen.

You're basically just advocating for us to put our heads in the sand.

I'll pass.

 

[Dr Pepper]

You're basically just advocating for us to put our heads in the sand.

I'll pass.

Not necessarily, it just gets tiresome when analytics and "5on5 xG/60" numbers get thrown around so frequently.